Every Number Here Ran on One RTX 3060
No cluster, no A100: the entire benchmark ran on one RTX 3060 with 12 GB. What fits on 12 GB, what to do with 8, and how to run with no local GPU at all through any OpenAI-compatible endpoint.
3 min read
Blog
Notes from the forest
Engineering notes on navigation, ingestion, agent memory and honest benchmarking, written by the people who ran the experiments.
Ingesting 100 Real Documents: The Write-Time Trade
Ingest is where the forest spends its effort: 1.71 s per document, a sixty-token summary contract passed at 100%, zero broken links. Pay once at write time, in a batch you can watch, instead of on every query forever.
4 min read
The Usual Patches, and Where They Stop
Rerankers, rewriting, HyDE, loops and agents all help, which is why they are everywhere. Here is what each one costs, where each one stops on multi-hop, and the gap that remains, measured.
4 min read
Run It Without a Local GPU
The benchmark rig is one RTX 3060, but nothing requires you to own a GPU. Point the chat binding at any OpenAI-compatible endpoint and let the forest do the structural work.
4 min read
Tokens per Correct Answer, Not per Query
Per-query cost makes failure look like a discount. Measured per correct answer, the forest navigator spends 0.58x the tokens of an iterative-RAG baseline, same 12B model. Here is the arithmetic.
4 min read
BM25 Over SQLite vs a Vector Database
The category assumes an embedding service, an ANN index and a GPU. We measured entry search at recall@5 = 1.00 at 1.3 ms p95 with BM25 over SQLite, and kept the embedder optional.
4 min read
Agent Amnesia and the Second Bill
Every agent session starts by rebuilding context the last one already paid for. What a persistent, growable forest memory changes about agent workflows, and the skills file that makes agents actually use it.
4 min read
The Forest Principle: spend intelligence on the environment, not the model
Stigmergy applied to a corpus. What a navigable environment concretely means, what it buys in tokens and hardware, and what it honestly does not fix.
3 min read
Want the long version?
The paper carries the full architecture, the benchmark tables and the findings that failed their criteria.